rlberry API¶

Manager¶

Main classes¶

`manager.ExperimentManager`(agent_class[, ...])	Class to train, optimize hyperparameters, evaluate and gather statistics about an agent.
`manager.MultipleManagers`([max_workers, ...])	Class to fit multiple ExperimentManager instances in parallel with multiple threads.

Evaluation and plot¶

manager.AdastopComparator([n, K, B, ...])

Compare sequentially agents, with possible early stopping.

`manager.evaluate_agents`(experiment_manager_list)	Evaluate and compare each of the agents in experiment_manager_list.
`manager.read_writer_data`(data_source[, ...])	Given a list of ExperimentManager or a folder, read data (corresponding to info) obtained in each episode.
`manager.plot_writer_data`(data_source, tag[, ...])	Given a list of ExperimentManager or a folder, plot data (corresponding to info) obtained in each episode.
`manager.plot_curves_smoothed_NW`(data, x, y)	Plot the performances contained in the data (see data parameter to learn what format it should be).
`manager.plot_curves_with_same_x`(data, x, y)	Plot the performances contained in the data (see data parameter to learn what format it should be).
`manager.compare_agents`(agent_source[, ...])	Compare several trained agents using the mean over n_simulations evaluations for each agent.
`manager.tensorboard_to_dataframe`(...)	Function to convert 'tensorboard log' to 'Panda DataFrames'.

Agents¶

Base classes¶

`agents.Agent`([env, eval_env, copy_env, ...])	Basic interface for agents.
`agents.AgentWithSimplePolicy`([env, ...])	Interface for agents whose policy is a function of observations only.

Agent importation tools¶

agents.stable_baselines.StableBaselinesAgent(env)

Wraps an StableBaselines3 Algorithm with a rlberry Agent.

Environments¶

Base class¶

`envs.interface.Model`()	Base class for an environment model.
`envs.basewrapper.Wrapper`(env[, wrap_spaces])	Wraps a given environment, similar to OpenAI gym's wrapper [1,2] (now updated to gymnasium).

Spaces¶

`spaces.Discrete`(n)	Class that represents discrete spaces.
`spaces.Box`(low, high[, shape, dtype])	Class that represents a space that is a cartesian product in R^n:
`spaces.Tuple`(spaces)	Inherited from gymnasium.spaces.Tuple for compatibility with gym.
`spaces.MultiDiscrete`(nvec[, dtype])	Inherited from gymnasium.spaces.MultiDiscrete for compatibility with gym.
`spaces.MultiBinary`(n)	Inherited from gymnasium.spaces.MultiBinary for compatibility with gym.
`spaces.Dict`([spaces])	Inherited from gymnasium.spaces.Dict for compatibility with gym.

Environment tools¶

`envs.gym_make`(id[, wrap_spaces])	Same as gym.make, but wraps the environment to ensure unified seeding with rlberry.
`envs.atari_make`(id[, seed])	Adaptor to manage Atari Env
`envs.PipelineEnv`(env_ctor, env_kwargs, wrappers)	Environment defined as a pipeline of wrappers and an environment to wrap.

Seeding¶

seeding.seeder.Seeder([seed_seq, spawn_seed_seq])

Base class to define objects that use random number generators.

`seeding.safe_reseed`(obj, seeder[, reseed_spaces])	Calls obj.reseed(seed_seq) method if available; If a obj.seed() method is available, call obj.seed(seed_val), where seed_val is generated by the seeder.
`seeding.set_external_seed`(seeder)	Set seeds of external libraries.

Utilities, Logging & Typing¶

Manager Utilitis¶

manager.preset_manager(*args, **kwds)

Preset an ExperimentManager to some fixed keywords.

Virtual environment Utilities¶

`manager.with_venv`([import_libs, ...])	Decorator to run the script in a function using a virtual environment.
`manager.run_venv_xp`([venv_dir_name, verbose])	Function to launch experiments in the current script file in virtual environments.

Writer Utilities¶

`utils.writers.DefaultWriter`(name[, ...])	Default writer to be used by the agents, optionally wraps an instance of tensorboard.SummaryWriter.
`agents.utils.replay.ReplayBuffer`(...[, ...])	Replay buffer that allows sampling data with shape (batch_size, time_size, ...).

Check Utilities¶

utils.check_rl_agent(agent[, env, init_kwargs])

Check ExperimentManager compatibility and check reproducibility/seeding.

utils.check_env(env)

Check that the environment is (almost) gym-compatible and that it is reproducible in the sense that it returns the same states when given the same seed.

utils.check_save_load(agent[, env, init_kwargs])

utils.check_fit_additive(agent[, env, ...])

Check that fitting two times with 10 fit budget is the same as fitting one time with 20 fit budget.

utils.check_seeding_agent(agent[, env, ...])

Check that the agent is reproducible.

utils.check_experiment_manager(agent[, env, ...])

Check that the agent is compatible with ExperimentManager.

Logging Utilities¶

utils.logging.set_level([level])

Set rlberry's logger level.

Environment Wrappers¶

`wrappers.discretize_state.DiscretizeStateWrapper`(...)	Discretize an environment with continuous states and discrete actions.
`wrappers.gym_utils.OldGymCompatibilityWrapper`(env)	Allow to use old gym env (V0.21) with rlberry (gymnasium).
`wrappers.RescaleRewardWrapper`(env, reward_range)	Rescale the reward function to a bounded range.

`utils.check_rl_agent`(agent[, env, init_kwargs])	Check ExperimentManager compatibility and check reproducibility/seeding.
`utils.check_env`(env)	Check that the environment is (almost) gym-compatible and that it is reproducible in the sense that it returns the same states when given the same seed.
`utils.check_save_load`(agent[, env, init_kwargs])
`utils.check_fit_additive`(agent[, env, ...])	Check that fitting two times with 10 fit budget is the same as fitting one time with 20 fit budget.
`utils.check_seeding_agent`(agent[, env, ...])	Check that the agent is reproducible.
`utils.check_experiment_manager`(agent[, env, ...])	Check that the agent is compatible with `ExperimentManager`.